xref: /freebsd/sys/vm/vm_object.c (revision 92f340d137ba5d6db7610ba1dae35842e2c9c8ea)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include "opt_vm.h"
66 
67 #include <sys/systm.h>
68 #include <sys/blockcount.h>
69 #include <sys/conf.h>
70 #include <sys/cpuset.h>
71 #include <sys/ipc.h>
72 #include <sys/jail.h>
73 #include <sys/limits.h>
74 #include <sys/lock.h>
75 #include <sys/mman.h>
76 #include <sys/mount.h>
77 #include <sys/kernel.h>
78 #include <sys/mutex.h>
79 #include <sys/pctrie.h>
80 #include <sys/proc.h>
81 #include <sys/refcount.h>
82 #include <sys/shm.h>
83 #include <sys/sx.h>
84 #include <sys/sysctl.h>
85 #include <sys/resourcevar.h>
86 #include <sys/refcount.h>
87 #include <sys/rwlock.h>
88 #include <sys/user.h>
89 #include <sys/vnode.h>
90 #include <sys/vmmeter.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_param.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/vm_phys.h>
101 #include <vm/vm_pagequeue.h>
102 #include <vm/swap_pager.h>
103 #include <vm/vm_kern.h>
104 #include <vm/vm_extern.h>
105 #include <vm/vm_radix.h>
106 #include <vm/vm_reserv.h>
107 #include <vm/uma.h>
108 
109 static int old_msync;
110 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
111     "Use old (insecure) msync behavior");
112 
113 static int	vm_object_page_collect_flush(struct pctrie_iter *pages,
114 		    vm_page_t p, int pagerflags, int flags, boolean_t *allclean,
115 		    boolean_t *eio);
116 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
117 		    boolean_t *allclean);
118 static void	vm_object_backing_remove(vm_object_t object);
119 
120 /*
121  *	Virtual memory objects maintain the actual data
122  *	associated with allocated virtual memory.  A given
123  *	page of memory exists within exactly one object.
124  *
125  *	An object is only deallocated when all "references"
126  *	are given up.  Only one "reference" to a given
127  *	region of an object should be writeable.
128  *
129  *	Associated with each object is a list of all resident
130  *	memory pages belonging to that object; this list is
131  *	maintained by the "vm_page" module, and locked by the object's
132  *	lock.
133  *
134  *	Each object also records a "pager" routine which is
135  *	used to retrieve (and store) pages to the proper backing
136  *	storage.  In addition, objects may be backed by other
137  *	objects from which they were virtual-copied.
138  *
139  *	The only items within the object structure which are
140  *	modified after time of creation are:
141  *		reference count		locked by object's lock
142  *		pager routine		locked by object's lock
143  *
144  */
145 
146 struct object_q vm_object_list;
147 struct mtx vm_object_list_mtx;	/* lock for object list and count */
148 
149 struct vm_object kernel_object_store;
150 
151 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
152     "VM object stats");
153 
154 static COUNTER_U64_DEFINE_EARLY(object_collapses);
155 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
156     &object_collapses,
157     "VM object collapses");
158 
159 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
160 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
161     &object_bypasses,
162     "VM object bypasses");
163 
164 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
165 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
166     &object_collapse_waits,
167     "Number of sleeps for collapse");
168 
169 static uma_zone_t obj_zone;
170 
171 static int vm_object_zinit(void *mem, int size, int flags);
172 
173 #ifdef INVARIANTS
174 static void vm_object_zdtor(void *mem, int size, void *arg);
175 
176 static void
177 vm_object_zdtor(void *mem, int size, void *arg)
178 {
179 	vm_object_t object;
180 
181 	object = (vm_object_t)mem;
182 	KASSERT(object->ref_count == 0,
183 	    ("object %p ref_count = %d", object, object->ref_count));
184 	KASSERT(TAILQ_EMPTY(&object->memq),
185 	    ("object %p has resident pages in its memq", object));
186 	KASSERT(vm_radix_is_empty(&object->rtree),
187 	    ("object %p has resident pages in its trie", object));
188 #if VM_NRESERVLEVEL > 0
189 	KASSERT(LIST_EMPTY(&object->rvq),
190 	    ("object %p has reservations",
191 	    object));
192 #endif
193 	KASSERT(!vm_object_busied(object),
194 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
195 	KASSERT(object->resident_page_count == 0,
196 	    ("object %p resident_page_count = %d",
197 	    object, object->resident_page_count));
198 	KASSERT(atomic_load_int(&object->shadow_count) == 0,
199 	    ("object %p shadow_count = %d",
200 	    object, atomic_load_int(&object->shadow_count)));
201 	KASSERT(object->type == OBJT_DEAD,
202 	    ("object %p has non-dead type %d",
203 	    object, object->type));
204 	KASSERT(object->charge == 0 && object->cred == NULL,
205 	    ("object %p has non-zero charge %ju (%p)",
206 	    object, (uintmax_t)object->charge, object->cred));
207 }
208 #endif
209 
210 static int
211 vm_object_zinit(void *mem, int size, int flags)
212 {
213 	vm_object_t object;
214 
215 	object = (vm_object_t)mem;
216 	rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW);
217 
218 	/* These are true for any object that has been freed */
219 	object->type = OBJT_DEAD;
220 	vm_radix_init(&object->rtree);
221 	refcount_init(&object->ref_count, 0);
222 	blockcount_init(&object->paging_in_progress);
223 	blockcount_init(&object->busy);
224 	object->resident_page_count = 0;
225 	atomic_store_int(&object->shadow_count, 0);
226 	object->flags = OBJ_DEAD;
227 
228 	mtx_lock(&vm_object_list_mtx);
229 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
230 	mtx_unlock(&vm_object_list_mtx);
231 	return (0);
232 }
233 
234 static void
235 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
236     vm_object_t object, void *handle)
237 {
238 
239 	TAILQ_INIT(&object->memq);
240 	LIST_INIT(&object->shadow_head);
241 
242 	object->type = type;
243 	object->flags = flags;
244 	if ((flags & OBJ_SWAP) != 0) {
245 		pctrie_init(&object->un_pager.swp.swp_blks);
246 		object->un_pager.swp.writemappings = 0;
247 	}
248 
249 	/*
250 	 * Ensure that swap_pager_swapoff() iteration over object_list
251 	 * sees up to date type and pctrie head if it observed
252 	 * non-dead object.
253 	 */
254 	atomic_thread_fence_rel();
255 
256 	object->pg_color = 0;
257 	object->size = size;
258 	object->domain.dr_policy = NULL;
259 	object->generation = 1;
260 	object->cleangeneration = 1;
261 	refcount_init(&object->ref_count, 1);
262 	object->memattr = VM_MEMATTR_DEFAULT;
263 	object->cred = NULL;
264 	object->charge = 0;
265 	object->handle = handle;
266 	object->backing_object = NULL;
267 	object->backing_object_offset = (vm_ooffset_t) 0;
268 #if VM_NRESERVLEVEL > 0
269 	LIST_INIT(&object->rvq);
270 #endif
271 	umtx_shm_object_init(object);
272 }
273 
274 /*
275  *	vm_object_init:
276  *
277  *	Initialize the VM objects module.
278  */
279 void
280 vm_object_init(void)
281 {
282 	TAILQ_INIT(&vm_object_list);
283 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
284 
285 	rw_init(&kernel_object->lock, "kernel vm object");
286 	vm_radix_init(&kernel_object->rtree);
287 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
288 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
289 #if VM_NRESERVLEVEL > 0
290 	kernel_object->flags |= OBJ_COLORED;
291 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
292 #endif
293 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
294 
295 	/*
296 	 * The lock portion of struct vm_object must be type stable due
297 	 * to vm_pageout_fallback_object_lock locking a vm object
298 	 * without holding any references to it.
299 	 *
300 	 * paging_in_progress is valid always.  Lockless references to
301 	 * the objects may acquire pip and then check OBJ_DEAD.
302 	 */
303 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
304 #ifdef INVARIANTS
305 	    vm_object_zdtor,
306 #else
307 	    NULL,
308 #endif
309 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
310 
311 	vm_radix_zinit();
312 }
313 
314 void
315 vm_object_clear_flag(vm_object_t object, u_short bits)
316 {
317 
318 	VM_OBJECT_ASSERT_WLOCKED(object);
319 	object->flags &= ~bits;
320 }
321 
322 /*
323  *	Sets the default memory attribute for the specified object.  Pages
324  *	that are allocated to this object are by default assigned this memory
325  *	attribute.
326  *
327  *	Presently, this function must be called before any pages are allocated
328  *	to the object.  In the future, this requirement may be relaxed for
329  *	"default" and "swap" objects.
330  */
331 int
332 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
333 {
334 
335 	VM_OBJECT_ASSERT_WLOCKED(object);
336 
337 	if (object->type == OBJT_DEAD)
338 		return (KERN_INVALID_ARGUMENT);
339 	if (!vm_radix_is_empty(&object->rtree))
340 		return (KERN_FAILURE);
341 
342 	object->memattr = memattr;
343 	return (KERN_SUCCESS);
344 }
345 
346 void
347 vm_object_pip_add(vm_object_t object, short i)
348 {
349 
350 	if (i > 0)
351 		blockcount_acquire(&object->paging_in_progress, i);
352 }
353 
354 void
355 vm_object_pip_wakeup(vm_object_t object)
356 {
357 
358 	vm_object_pip_wakeupn(object, 1);
359 }
360 
361 void
362 vm_object_pip_wakeupn(vm_object_t object, short i)
363 {
364 
365 	if (i > 0)
366 		blockcount_release(&object->paging_in_progress, i);
367 }
368 
369 /*
370  * Atomically drop the object lock and wait for pip to drain.  This protects
371  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
372  * on return.
373  */
374 static void
375 vm_object_pip_sleep(vm_object_t object, const char *waitid)
376 {
377 
378 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
379 	    waitid, PVM | PDROP);
380 }
381 
382 void
383 vm_object_pip_wait(vm_object_t object, const char *waitid)
384 {
385 
386 	VM_OBJECT_ASSERT_WLOCKED(object);
387 
388 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
389 	    PVM);
390 }
391 
392 void
393 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
394 {
395 
396 	VM_OBJECT_ASSERT_UNLOCKED(object);
397 
398 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
399 }
400 
401 /*
402  *	vm_object_allocate:
403  *
404  *	Returns a new object with the given size.
405  */
406 vm_object_t
407 vm_object_allocate(objtype_t type, vm_pindex_t size)
408 {
409 	vm_object_t object;
410 	u_short flags;
411 
412 	switch (type) {
413 	case OBJT_DEAD:
414 		panic("vm_object_allocate: can't create OBJT_DEAD");
415 	case OBJT_SWAP:
416 		flags = OBJ_COLORED | OBJ_SWAP;
417 		break;
418 	case OBJT_DEVICE:
419 	case OBJT_SG:
420 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
421 		break;
422 	case OBJT_MGTDEVICE:
423 		flags = OBJ_FICTITIOUS;
424 		break;
425 	case OBJT_PHYS:
426 		flags = OBJ_UNMANAGED;
427 		break;
428 	case OBJT_VNODE:
429 		flags = 0;
430 		break;
431 	default:
432 		panic("vm_object_allocate: type %d is undefined or dynamic",
433 		    type);
434 	}
435 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
436 	_vm_object_allocate(type, size, flags, object, NULL);
437 
438 	return (object);
439 }
440 
441 vm_object_t
442 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
443 {
444 	vm_object_t object;
445 
446 	MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
447 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
448 	_vm_object_allocate(dyntype, size, flags, object, NULL);
449 
450 	return (object);
451 }
452 
453 /*
454  *	vm_object_allocate_anon:
455  *
456  *	Returns a new default object of the given size and marked as
457  *	anonymous memory for special split/collapse handling.  Color
458  *	to be initialized by the caller.
459  */
460 vm_object_t
461 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
462     struct ucred *cred, vm_size_t charge)
463 {
464 	vm_object_t handle, object;
465 
466 	if (backing_object == NULL)
467 		handle = NULL;
468 	else if ((backing_object->flags & OBJ_ANON) != 0)
469 		handle = backing_object->handle;
470 	else
471 		handle = backing_object;
472 	object = uma_zalloc(obj_zone, M_WAITOK);
473 	_vm_object_allocate(OBJT_SWAP, size,
474 	    OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle);
475 	object->cred = cred;
476 	object->charge = cred != NULL ? charge : 0;
477 	return (object);
478 }
479 
480 static void
481 vm_object_reference_vnode(vm_object_t object)
482 {
483 	u_int old;
484 
485 	/*
486 	 * vnode objects need the lock for the first reference
487 	 * to serialize with vnode_object_deallocate().
488 	 */
489 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
490 		VM_OBJECT_RLOCK(object);
491 		old = refcount_acquire(&object->ref_count);
492 		if (object->type == OBJT_VNODE && old == 0)
493 			vref(object->handle);
494 		VM_OBJECT_RUNLOCK(object);
495 	}
496 }
497 
498 /*
499  *	vm_object_reference:
500  *
501  *	Acquires a reference to the given object.
502  */
503 void
504 vm_object_reference(vm_object_t object)
505 {
506 
507 	if (object == NULL)
508 		return;
509 
510 	if (object->type == OBJT_VNODE)
511 		vm_object_reference_vnode(object);
512 	else
513 		refcount_acquire(&object->ref_count);
514 	KASSERT((object->flags & OBJ_DEAD) == 0,
515 	    ("vm_object_reference: Referenced dead object."));
516 }
517 
518 /*
519  *	vm_object_reference_locked:
520  *
521  *	Gets another reference to the given object.
522  *
523  *	The object must be locked.
524  */
525 void
526 vm_object_reference_locked(vm_object_t object)
527 {
528 	u_int old;
529 
530 	VM_OBJECT_ASSERT_LOCKED(object);
531 	old = refcount_acquire(&object->ref_count);
532 	if (object->type == OBJT_VNODE && old == 0)
533 		vref(object->handle);
534 	KASSERT((object->flags & OBJ_DEAD) == 0,
535 	    ("vm_object_reference: Referenced dead object."));
536 }
537 
538 /*
539  * Handle deallocating an object of type OBJT_VNODE.
540  */
541 static void
542 vm_object_deallocate_vnode(vm_object_t object)
543 {
544 	struct vnode *vp = (struct vnode *) object->handle;
545 	bool last;
546 
547 	KASSERT(object->type == OBJT_VNODE,
548 	    ("vm_object_deallocate_vnode: not a vnode object"));
549 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
550 
551 	/* Object lock to protect handle lookup. */
552 	last = refcount_release(&object->ref_count);
553 	VM_OBJECT_RUNLOCK(object);
554 
555 	if (!last)
556 		return;
557 
558 	if (!umtx_shm_vnobj_persistent)
559 		umtx_shm_object_terminated(object);
560 
561 	/* vrele may need the vnode lock. */
562 	vrele(vp);
563 }
564 
565 /*
566  * We dropped a reference on an object and discovered that it had a
567  * single remaining shadow.  This is a sibling of the reference we
568  * dropped.  Attempt to collapse the sibling and backing object.
569  */
570 static vm_object_t
571 vm_object_deallocate_anon(vm_object_t backing_object)
572 {
573 	vm_object_t object;
574 
575 	/* Fetch the final shadow.  */
576 	object = LIST_FIRST(&backing_object->shadow_head);
577 	KASSERT(object != NULL &&
578 	    atomic_load_int(&backing_object->shadow_count) == 1,
579 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
580 	    backing_object->ref_count,
581 	    atomic_load_int(&backing_object->shadow_count)));
582 	KASSERT((object->flags & OBJ_ANON) != 0,
583 	    ("invalid shadow object %p", object));
584 
585 	if (!VM_OBJECT_TRYWLOCK(object)) {
586 		/*
587 		 * Prevent object from disappearing since we do not have a
588 		 * reference.
589 		 */
590 		vm_object_pip_add(object, 1);
591 		VM_OBJECT_WUNLOCK(backing_object);
592 		VM_OBJECT_WLOCK(object);
593 		vm_object_pip_wakeup(object);
594 	} else
595 		VM_OBJECT_WUNLOCK(backing_object);
596 
597 	/*
598 	 * Check for a collapse/terminate race with the last reference holder.
599 	 */
600 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
601 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
602 		VM_OBJECT_WUNLOCK(object);
603 		return (NULL);
604 	}
605 	backing_object = object->backing_object;
606 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
607 		vm_object_collapse(object);
608 	VM_OBJECT_WUNLOCK(object);
609 
610 	return (object);
611 }
612 
613 /*
614  *	vm_object_deallocate:
615  *
616  *	Release a reference to the specified object,
617  *	gained either through a vm_object_allocate
618  *	or a vm_object_reference call.  When all references
619  *	are gone, storage associated with this object
620  *	may be relinquished.
621  *
622  *	No object may be locked.
623  */
624 void
625 vm_object_deallocate(vm_object_t object)
626 {
627 	vm_object_t temp;
628 	bool released;
629 
630 	while (object != NULL) {
631 		/*
632 		 * If the reference count goes to 0 we start calling
633 		 * vm_object_terminate() on the object chain.  A ref count
634 		 * of 1 may be a special case depending on the shadow count
635 		 * being 0 or 1.  These cases require a write lock on the
636 		 * object.
637 		 */
638 		if ((object->flags & OBJ_ANON) == 0)
639 			released = refcount_release_if_gt(&object->ref_count, 1);
640 		else
641 			released = refcount_release_if_gt(&object->ref_count, 2);
642 		if (released)
643 			return;
644 
645 		if (object->type == OBJT_VNODE) {
646 			VM_OBJECT_RLOCK(object);
647 			if (object->type == OBJT_VNODE) {
648 				vm_object_deallocate_vnode(object);
649 				return;
650 			}
651 			VM_OBJECT_RUNLOCK(object);
652 		}
653 
654 		VM_OBJECT_WLOCK(object);
655 		KASSERT(object->ref_count > 0,
656 		    ("vm_object_deallocate: object deallocated too many times: %d",
657 		    object->type));
658 
659 		/*
660 		 * If this is not the final reference to an anonymous
661 		 * object we may need to collapse the shadow chain.
662 		 */
663 		if (!refcount_release(&object->ref_count)) {
664 			if (object->ref_count > 1 ||
665 			    atomic_load_int(&object->shadow_count) == 0) {
666 				if ((object->flags & OBJ_ANON) != 0 &&
667 				    object->ref_count == 1)
668 					vm_object_set_flag(object,
669 					    OBJ_ONEMAPPING);
670 				VM_OBJECT_WUNLOCK(object);
671 				return;
672 			}
673 
674 			/* Handle collapsing last ref on anonymous objects. */
675 			object = vm_object_deallocate_anon(object);
676 			continue;
677 		}
678 
679 		/*
680 		 * Handle the final reference to an object.  We restart
681 		 * the loop with the backing object to avoid recursion.
682 		 */
683 		umtx_shm_object_terminated(object);
684 		temp = object->backing_object;
685 		if (temp != NULL) {
686 			KASSERT(object->type == OBJT_SWAP,
687 			    ("shadowed tmpfs v_object 2 %p", object));
688 			vm_object_backing_remove(object);
689 		}
690 
691 		KASSERT((object->flags & OBJ_DEAD) == 0,
692 		    ("vm_object_deallocate: Terminating dead object."));
693 		vm_object_set_flag(object, OBJ_DEAD);
694 		vm_object_terminate(object);
695 		object = temp;
696 	}
697 }
698 
699 void
700 vm_object_destroy(vm_object_t object)
701 {
702 	uma_zfree(obj_zone, object);
703 }
704 
705 static void
706 vm_object_sub_shadow(vm_object_t object)
707 {
708 	KASSERT(object->shadow_count >= 1,
709 	    ("object %p sub_shadow count zero", object));
710 	atomic_subtract_int(&object->shadow_count, 1);
711 }
712 
713 static void
714 vm_object_backing_remove_locked(vm_object_t object)
715 {
716 	vm_object_t backing_object;
717 
718 	backing_object = object->backing_object;
719 	VM_OBJECT_ASSERT_WLOCKED(object);
720 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
721 
722 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
723 	    ("vm_object_backing_remove: Removing collapsing object."));
724 
725 	vm_object_sub_shadow(backing_object);
726 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
727 		LIST_REMOVE(object, shadow_list);
728 		vm_object_clear_flag(object, OBJ_SHADOWLIST);
729 	}
730 	object->backing_object = NULL;
731 }
732 
733 static void
734 vm_object_backing_remove(vm_object_t object)
735 {
736 	vm_object_t backing_object;
737 
738 	VM_OBJECT_ASSERT_WLOCKED(object);
739 
740 	backing_object = object->backing_object;
741 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
742 		VM_OBJECT_WLOCK(backing_object);
743 		vm_object_backing_remove_locked(object);
744 		VM_OBJECT_WUNLOCK(backing_object);
745 	} else {
746 		object->backing_object = NULL;
747 		vm_object_sub_shadow(backing_object);
748 	}
749 }
750 
751 static void
752 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
753 {
754 
755 	VM_OBJECT_ASSERT_WLOCKED(object);
756 
757 	atomic_add_int(&backing_object->shadow_count, 1);
758 	if ((backing_object->flags & OBJ_ANON) != 0) {
759 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
760 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
761 		    shadow_list);
762 		vm_object_set_flag(object, OBJ_SHADOWLIST);
763 	}
764 	object->backing_object = backing_object;
765 }
766 
767 static void
768 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
769 {
770 
771 	VM_OBJECT_ASSERT_WLOCKED(object);
772 
773 	if ((backing_object->flags & OBJ_ANON) != 0) {
774 		VM_OBJECT_WLOCK(backing_object);
775 		vm_object_backing_insert_locked(object, backing_object);
776 		VM_OBJECT_WUNLOCK(backing_object);
777 	} else {
778 		object->backing_object = backing_object;
779 		atomic_add_int(&backing_object->shadow_count, 1);
780 	}
781 }
782 
783 /*
784  * Insert an object into a backing_object's shadow list with an additional
785  * reference to the backing_object added.
786  */
787 static void
788 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
789 {
790 
791 	VM_OBJECT_ASSERT_WLOCKED(object);
792 
793 	if ((backing_object->flags & OBJ_ANON) != 0) {
794 		VM_OBJECT_WLOCK(backing_object);
795 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
796 		    ("shadowing dead anonymous object"));
797 		vm_object_reference_locked(backing_object);
798 		vm_object_backing_insert_locked(object, backing_object);
799 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
800 		VM_OBJECT_WUNLOCK(backing_object);
801 	} else {
802 		vm_object_reference(backing_object);
803 		atomic_add_int(&backing_object->shadow_count, 1);
804 		object->backing_object = backing_object;
805 	}
806 }
807 
808 /*
809  * Transfer a backing reference from backing_object to object.
810  */
811 static void
812 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
813 {
814 	vm_object_t new_backing_object;
815 
816 	/*
817 	 * Note that the reference to backing_object->backing_object
818 	 * moves from within backing_object to within object.
819 	 */
820 	vm_object_backing_remove_locked(object);
821 	new_backing_object = backing_object->backing_object;
822 	if (new_backing_object == NULL)
823 		return;
824 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
825 		VM_OBJECT_WLOCK(new_backing_object);
826 		vm_object_backing_remove_locked(backing_object);
827 		vm_object_backing_insert_locked(object, new_backing_object);
828 		VM_OBJECT_WUNLOCK(new_backing_object);
829 	} else {
830 		/*
831 		 * shadow_count for new_backing_object is left
832 		 * unchanged, its reference provided by backing_object
833 		 * is replaced by object.
834 		 */
835 		object->backing_object = new_backing_object;
836 		backing_object->backing_object = NULL;
837 	}
838 }
839 
840 /*
841  * Wait for a concurrent collapse to settle.
842  */
843 static void
844 vm_object_collapse_wait(vm_object_t object)
845 {
846 
847 	VM_OBJECT_ASSERT_WLOCKED(object);
848 
849 	while ((object->flags & OBJ_COLLAPSING) != 0) {
850 		vm_object_pip_wait(object, "vmcolwait");
851 		counter_u64_add(object_collapse_waits, 1);
852 	}
853 }
854 
855 /*
856  * Waits for a backing object to clear a pending collapse and returns
857  * it locked if it is an ANON object.
858  */
859 static vm_object_t
860 vm_object_backing_collapse_wait(vm_object_t object)
861 {
862 	vm_object_t backing_object;
863 
864 	VM_OBJECT_ASSERT_WLOCKED(object);
865 
866 	for (;;) {
867 		backing_object = object->backing_object;
868 		if (backing_object == NULL ||
869 		    (backing_object->flags & OBJ_ANON) == 0)
870 			return (NULL);
871 		VM_OBJECT_WLOCK(backing_object);
872 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
873 			break;
874 		VM_OBJECT_WUNLOCK(object);
875 		vm_object_pip_sleep(backing_object, "vmbckwait");
876 		counter_u64_add(object_collapse_waits, 1);
877 		VM_OBJECT_WLOCK(object);
878 	}
879 	return (backing_object);
880 }
881 
882 /*
883  *	vm_object_terminate_single_page removes a pageable page from the object,
884  *	and removes it from the paging queues and frees it, if it is not wired.
885  *	It is invoked via callback from vm_object_terminate_pages.
886  */
887 static void
888 vm_object_terminate_single_page(vm_page_t p, void *objectv)
889 {
890 	vm_object_t object __diagused = objectv;
891 
892 	vm_page_assert_unbusied(p);
893 	KASSERT(p->object == object &&
894 	    (p->ref_count & VPRC_OBJREF) != 0,
895 	    ("%s: page %p is inconsistent", __func__, p));
896 	p->object = NULL;
897 	if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
898 		KASSERT((object->flags & OBJ_UNMANAGED) != 0 ||
899 		    vm_page_astate_load(p).queue != PQ_NONE,
900 		    ("%s: page %p does not belong to a queue", __func__, p));
901 		VM_CNT_INC(v_pfree);
902 		vm_page_free(p);
903 	}
904 }
905 
906 /*
907  *	vm_object_terminate_pages removes any remaining pageable pages
908  *	from the object and resets the object to an empty state.
909  */
910 static void
911 vm_object_terminate_pages(vm_object_t object)
912 {
913 	VM_OBJECT_ASSERT_WLOCKED(object);
914 
915 	/*
916 	 * If the object contained any pages, then reset it to an empty state.
917 	 * Rather than incrementally removing each page from the object, the
918 	 * page and object are reset to any empty state.
919 	 */
920 	if (object->resident_page_count == 0)
921 		return;
922 
923 	vm_radix_reclaim_callback(&object->rtree,
924 	    vm_object_terminate_single_page, object);
925 	TAILQ_INIT(&object->memq);
926 	object->resident_page_count = 0;
927 	if (object->type == OBJT_VNODE)
928 		vdrop(object->handle);
929 }
930 
931 /*
932  *	vm_object_terminate actually destroys the specified object, freeing
933  *	up all previously used resources.
934  *
935  *	The object must be locked.
936  *	This routine may block.
937  */
938 void
939 vm_object_terminate(vm_object_t object)
940 {
941 
942 	VM_OBJECT_ASSERT_WLOCKED(object);
943 	KASSERT((object->flags & OBJ_DEAD) != 0,
944 	    ("terminating non-dead obj %p", object));
945 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
946 	    ("terminating collapsing obj %p", object));
947 	KASSERT(object->backing_object == NULL,
948 	    ("terminating shadow obj %p", object));
949 
950 	/*
951 	 * Wait for the pageout daemon and other current users to be
952 	 * done with the object.  Note that new paging_in_progress
953 	 * users can come after this wait, but they must check
954 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
955 	 * the object being terminated.
956 	 */
957 	vm_object_pip_wait(object, "objtrm");
958 
959 	KASSERT(object->ref_count == 0,
960 	    ("vm_object_terminate: object with references, ref_count=%d",
961 	    object->ref_count));
962 
963 	if ((object->flags & OBJ_PG_DTOR) == 0)
964 		vm_object_terminate_pages(object);
965 
966 #if VM_NRESERVLEVEL > 0
967 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
968 		vm_reserv_break_all(object);
969 #endif
970 
971 	KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0,
972 	    ("%s: non-swap obj %p has cred", __func__, object));
973 
974 	/*
975 	 * Let the pager know object is dead.
976 	 */
977 	vm_pager_deallocate(object);
978 	VM_OBJECT_WUNLOCK(object);
979 
980 	vm_object_destroy(object);
981 }
982 
983 /*
984  * Make the page read-only so that we can clear the object flags.  However, if
985  * this is a nosync mmap then the object is likely to stay dirty so do not
986  * mess with the page and do not clear the object flags.  Returns TRUE if the
987  * page should be flushed, and FALSE otherwise.
988  */
989 static boolean_t
990 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
991 {
992 
993 	vm_page_assert_busied(p);
994 
995 	/*
996 	 * If we have been asked to skip nosync pages and this is a
997 	 * nosync page, skip it.  Note that the object flags were not
998 	 * cleared in this case so we do not have to set them.
999 	 */
1000 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
1001 		*allclean = FALSE;
1002 		return (FALSE);
1003 	} else {
1004 		pmap_remove_write(p);
1005 		return (p->dirty != 0);
1006 	}
1007 }
1008 
1009 /*
1010  *	vm_object_page_clean
1011  *
1012  *	Clean all dirty pages in the specified range of object.  Leaves page
1013  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1014  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1015  *	leaving the object dirty.
1016  *
1017  *	For swap objects backing tmpfs regular files, do not flush anything,
1018  *	but remove write protection on the mapped pages to update mtime through
1019  *	mmaped writes.
1020  *
1021  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1022  *	synchronous clustering mode implementation.
1023  *
1024  *	Odd semantics: if start == end, we clean everything.
1025  *
1026  *	The object must be locked.
1027  *
1028  *	Returns FALSE if some page from the range was not written, as
1029  *	reported by the pager, and TRUE otherwise.
1030  */
1031 boolean_t
1032 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1033     int flags)
1034 {
1035 	struct pctrie_iter pages;
1036 	vm_page_t np, p;
1037 	vm_pindex_t pi, tend, tstart;
1038 	int curgeneration, n, pagerflags;
1039 	boolean_t eio, res, allclean;
1040 
1041 	VM_OBJECT_ASSERT_WLOCKED(object);
1042 
1043 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1044 		return (TRUE);
1045 
1046 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1047 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1048 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1049 
1050 	tstart = OFF_TO_IDX(start);
1051 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1052 	allclean = tstart == 0 && tend >= object->size;
1053 	res = TRUE;
1054 	vm_page_iter_init(&pages, object);
1055 
1056 rescan:
1057 	curgeneration = object->generation;
1058 
1059 	for (p = vm_radix_iter_lookup_ge(&pages, tstart); p != NULL; p = np) {
1060 		pi = p->pindex;
1061 		if (pi >= tend)
1062 			break;
1063 		if (vm_page_none_valid(p)) {
1064 			np = vm_radix_iter_step(&pages);
1065 			continue;
1066 		}
1067 		if (!vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL)) {
1068 			pctrie_iter_reset(&pages);
1069 			if (object->generation != curgeneration &&
1070 			    (flags & OBJPC_SYNC) != 0)
1071 				goto rescan;
1072 			np = vm_radix_iter_lookup_ge(&pages, pi);
1073 			continue;
1074 		}
1075 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1076 			np = vm_radix_iter_step(&pages);
1077 			vm_page_xunbusy(p);
1078 			continue;
1079 		}
1080 		if (object->type == OBJT_VNODE) {
1081 			n = vm_object_page_collect_flush(&pages, p, pagerflags,
1082 			    flags, &allclean, &eio);
1083 			pctrie_iter_reset(&pages);
1084 			if (eio) {
1085 				res = FALSE;
1086 				allclean = FALSE;
1087 			}
1088 			if (object->generation != curgeneration &&
1089 			    (flags & OBJPC_SYNC) != 0)
1090 				goto rescan;
1091 
1092 			/*
1093 			 * If the VOP_PUTPAGES() did a truncated write, so
1094 			 * that even the first page of the run is not fully
1095 			 * written, vm_pageout_flush() returns 0 as the run
1096 			 * length.  Since the condition that caused truncated
1097 			 * write may be permanent, e.g. exhausted free space,
1098 			 * accepting n == 0 would cause an infinite loop.
1099 			 *
1100 			 * Forwarding the iterator leaves the unwritten page
1101 			 * behind, but there is not much we can do there if
1102 			 * filesystem refuses to write it.
1103 			 */
1104 			if (n == 0) {
1105 				n = 1;
1106 				allclean = FALSE;
1107 			}
1108 		} else {
1109 			n = 1;
1110 			vm_page_xunbusy(p);
1111 		}
1112 		np = vm_radix_iter_lookup_ge(&pages, pi + n);
1113 	}
1114 #if 0
1115 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1116 #endif
1117 
1118 	/*
1119 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1120 	 * scan.  It needs to update mtime, which happens for other
1121 	 * filesystems during page writeouts.
1122 	 */
1123 	if (allclean && object->type == OBJT_VNODE)
1124 		object->cleangeneration = curgeneration;
1125 	return (res);
1126 }
1127 
1128 static int
1129 vm_object_page_collect_flush(struct pctrie_iter *pages, vm_page_t p,
1130     int pagerflags, int flags, boolean_t *allclean, boolean_t *eio)
1131 {
1132 	vm_page_t ma[2 * vm_pageout_page_count - 1];
1133 	int base, count, runlen;
1134 
1135 	vm_page_lock_assert(p, MA_NOTOWNED);
1136 	vm_page_assert_xbusied(p);
1137 	base = nitems(ma) / 2;
1138 	ma[base] = p;
1139 	for (count = 1; count < vm_pageout_page_count; count++) {
1140 		p = vm_radix_iter_next(pages);
1141 		if (p == NULL || vm_page_tryxbusy(p) == 0)
1142 			break;
1143 		if (!vm_object_page_remove_write(p, flags, allclean)) {
1144 			vm_page_xunbusy(p);
1145 			break;
1146 		}
1147 		ma[base + count] = p;
1148 	}
1149 
1150 	pages->index = ma[base]->pindex;
1151 	for (; count < vm_pageout_page_count; count++) {
1152 		p = vm_radix_iter_prev(pages);
1153 		if (p == NULL || vm_page_tryxbusy(p) == 0)
1154 			break;
1155 		if (!vm_object_page_remove_write(p, flags, allclean)) {
1156 			vm_page_xunbusy(p);
1157 			break;
1158 		}
1159 		ma[--base] = p;
1160 	}
1161 
1162 	vm_pageout_flush(&ma[base], count, pagerflags, nitems(ma) / 2 - base,
1163 	    &runlen, eio);
1164 	return (runlen);
1165 }
1166 
1167 /*
1168  * Note that there is absolutely no sense in writing out
1169  * anonymous objects, so we track down the vnode object
1170  * to write out.
1171  * We invalidate (remove) all pages from the address space
1172  * for semantic correctness.
1173  *
1174  * If the backing object is a device object with unmanaged pages, then any
1175  * mappings to the specified range of pages must be removed before this
1176  * function is called.
1177  *
1178  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1179  * may start out with a NULL object.
1180  */
1181 boolean_t
1182 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1183     boolean_t syncio, boolean_t invalidate)
1184 {
1185 	vm_object_t backing_object;
1186 	struct vnode *vp;
1187 	struct mount *mp;
1188 	int error, flags, fsync_after;
1189 	boolean_t res;
1190 
1191 	if (object == NULL)
1192 		return (TRUE);
1193 	res = TRUE;
1194 	error = 0;
1195 	VM_OBJECT_WLOCK(object);
1196 	while ((backing_object = object->backing_object) != NULL) {
1197 		VM_OBJECT_WLOCK(backing_object);
1198 		offset += object->backing_object_offset;
1199 		VM_OBJECT_WUNLOCK(object);
1200 		object = backing_object;
1201 		if (object->size < OFF_TO_IDX(offset + size))
1202 			size = IDX_TO_OFF(object->size) - offset;
1203 	}
1204 	/*
1205 	 * Flush pages if writing is allowed, invalidate them
1206 	 * if invalidation requested.  Pages undergoing I/O
1207 	 * will be ignored by vm_object_page_remove().
1208 	 *
1209 	 * We cannot lock the vnode and then wait for paging
1210 	 * to complete without deadlocking against vm_fault.
1211 	 * Instead we simply call vm_object_page_remove() and
1212 	 * allow it to block internally on a page-by-page
1213 	 * basis when it encounters pages undergoing async
1214 	 * I/O.
1215 	 */
1216 	if (object->type == OBJT_VNODE &&
1217 	    vm_object_mightbedirty(object) != 0 &&
1218 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1219 		VM_OBJECT_WUNLOCK(object);
1220 		(void)vn_start_write(vp, &mp, V_WAIT);
1221 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1222 		if (syncio && !invalidate && offset == 0 &&
1223 		    atop(size) == object->size) {
1224 			/*
1225 			 * If syncing the whole mapping of the file,
1226 			 * it is faster to schedule all the writes in
1227 			 * async mode, also allowing the clustering,
1228 			 * and then wait for i/o to complete.
1229 			 */
1230 			flags = 0;
1231 			fsync_after = TRUE;
1232 		} else {
1233 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1234 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1235 			fsync_after = FALSE;
1236 		}
1237 		VM_OBJECT_WLOCK(object);
1238 		res = vm_object_page_clean(object, offset, offset + size,
1239 		    flags);
1240 		VM_OBJECT_WUNLOCK(object);
1241 		if (fsync_after) {
1242 			for (;;) {
1243 				error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1244 				if (error != ERELOOKUP)
1245 					break;
1246 
1247 				/*
1248 				 * Allow SU/bufdaemon to handle more
1249 				 * dependencies in the meantime.
1250 				 */
1251 				VOP_UNLOCK(vp);
1252 				vn_finished_write(mp);
1253 
1254 				(void)vn_start_write(vp, &mp, V_WAIT);
1255 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1256 			}
1257 		}
1258 		VOP_UNLOCK(vp);
1259 		vn_finished_write(mp);
1260 		if (error != 0)
1261 			res = FALSE;
1262 		VM_OBJECT_WLOCK(object);
1263 	}
1264 	if ((object->type == OBJT_VNODE ||
1265 	     object->type == OBJT_DEVICE) && invalidate) {
1266 		if (object->type == OBJT_DEVICE)
1267 			/*
1268 			 * The option OBJPR_NOTMAPPED must be passed here
1269 			 * because vm_object_page_remove() cannot remove
1270 			 * unmanaged mappings.
1271 			 */
1272 			flags = OBJPR_NOTMAPPED;
1273 		else if (old_msync)
1274 			flags = 0;
1275 		else
1276 			flags = OBJPR_CLEANONLY;
1277 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1278 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1279 	}
1280 	VM_OBJECT_WUNLOCK(object);
1281 	return (res);
1282 }
1283 
1284 /*
1285  * Determine whether the given advice can be applied to the object.  Advice is
1286  * not applied to unmanaged pages since they never belong to page queues, and
1287  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1288  * have been mapped at most once.
1289  */
1290 static bool
1291 vm_object_advice_applies(vm_object_t object, int advice)
1292 {
1293 
1294 	if ((object->flags & OBJ_UNMANAGED) != 0)
1295 		return (false);
1296 	if (advice != MADV_FREE)
1297 		return (true);
1298 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1299 	    (OBJ_ONEMAPPING | OBJ_ANON));
1300 }
1301 
1302 static void
1303 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1304     vm_size_t size)
1305 {
1306 
1307 	if (advice == MADV_FREE)
1308 		vm_pager_freespace(object, pindex, size);
1309 }
1310 
1311 /*
1312  *	vm_object_madvise:
1313  *
1314  *	Implements the madvise function at the object/page level.
1315  *
1316  *	MADV_WILLNEED	(any object)
1317  *
1318  *	    Activate the specified pages if they are resident.
1319  *
1320  *	MADV_DONTNEED	(any object)
1321  *
1322  *	    Deactivate the specified pages if they are resident.
1323  *
1324  *	MADV_FREE	(OBJT_SWAP objects, OBJ_ONEMAPPING only)
1325  *
1326  *	    Deactivate and clean the specified pages if they are
1327  *	    resident.  This permits the process to reuse the pages
1328  *	    without faulting or the kernel to reclaim the pages
1329  *	    without I/O.
1330  */
1331 void
1332 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1333     int advice)
1334 {
1335 	struct pctrie_iter pages;
1336 	vm_pindex_t tpindex;
1337 	vm_object_t backing_object, tobject;
1338 	vm_page_t m, tm;
1339 
1340 	if (object == NULL)
1341 		return;
1342 
1343 	vm_page_iter_init(&pages, object);
1344 relookup:
1345 	VM_OBJECT_WLOCK(object);
1346 	if (!vm_object_advice_applies(object, advice)) {
1347 		VM_OBJECT_WUNLOCK(object);
1348 		return;
1349 	}
1350 	for (m = vm_radix_iter_lookup_ge(&pages, pindex); pindex < end;
1351 	    pindex++) {
1352 		tobject = object;
1353 
1354 		/*
1355 		 * If the next page isn't resident in the top-level object, we
1356 		 * need to search the shadow chain.  When applying MADV_FREE, we
1357 		 * take care to release any swap space used to store
1358 		 * non-resident pages.
1359 		 */
1360 		if (m == NULL || pindex < m->pindex) {
1361 			/*
1362 			 * Optimize a common case: if the top-level object has
1363 			 * no backing object, we can skip over the non-resident
1364 			 * range in constant time.
1365 			 */
1366 			if (object->backing_object == NULL) {
1367 				tpindex = (m != NULL && m->pindex < end) ?
1368 				    m->pindex : end;
1369 				vm_object_madvise_freespace(object, advice,
1370 				    pindex, tpindex - pindex);
1371 				if ((pindex = tpindex) == end)
1372 					break;
1373 				goto next_page;
1374 			}
1375 
1376 			tpindex = pindex;
1377 			do {
1378 				vm_object_madvise_freespace(tobject, advice,
1379 				    tpindex, 1);
1380 				/*
1381 				 * Prepare to search the next object in the
1382 				 * chain.
1383 				 */
1384 				backing_object = tobject->backing_object;
1385 				if (backing_object == NULL)
1386 					goto next_pindex;
1387 				VM_OBJECT_WLOCK(backing_object);
1388 				tpindex +=
1389 				    OFF_TO_IDX(tobject->backing_object_offset);
1390 				if (tobject != object)
1391 					VM_OBJECT_WUNLOCK(tobject);
1392 				tobject = backing_object;
1393 				if (!vm_object_advice_applies(tobject, advice))
1394 					goto next_pindex;
1395 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1396 			    NULL);
1397 		} else {
1398 next_page:
1399 			tm = m;
1400 			m = vm_radix_iter_step(&pages);
1401 		}
1402 
1403 		/*
1404 		 * If the page is not in a normal state, skip it.  The page
1405 		 * can not be invalidated while the object lock is held.
1406 		 */
1407 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1408 			goto next_pindex;
1409 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1410 		    ("vm_object_madvise: page %p is fictitious", tm));
1411 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1412 		    ("vm_object_madvise: page %p is not managed", tm));
1413 		if (vm_page_tryxbusy(tm) == 0) {
1414 			if (object != tobject)
1415 				VM_OBJECT_WUNLOCK(object);
1416 			if (advice == MADV_WILLNEED) {
1417 				/*
1418 				 * Reference the page before unlocking and
1419 				 * sleeping so that the page daemon is less
1420 				 * likely to reclaim it.
1421 				 */
1422 				vm_page_aflag_set(tm, PGA_REFERENCED);
1423 			}
1424 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1425 				VM_OBJECT_WUNLOCK(tobject);
1426 			pctrie_iter_reset(&pages);
1427   			goto relookup;
1428 		}
1429 		vm_page_advise(tm, advice);
1430 		vm_page_xunbusy(tm);
1431 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1432 next_pindex:
1433 		if (tobject != object)
1434 			VM_OBJECT_WUNLOCK(tobject);
1435 	}
1436 	VM_OBJECT_WUNLOCK(object);
1437 }
1438 
1439 /*
1440  *	vm_object_shadow:
1441  *
1442  *	Create a new object which is backed by the
1443  *	specified existing object range.  The source
1444  *	object reference is deallocated.
1445  *
1446  *	The new object and offset into that object
1447  *	are returned in the source parameters.
1448  */
1449 void
1450 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1451     struct ucred *cred, bool shared)
1452 {
1453 	vm_object_t source;
1454 	vm_object_t result;
1455 
1456 	source = *object;
1457 
1458 	/*
1459 	 * Don't create the new object if the old object isn't shared.
1460 	 *
1461 	 * If we hold the only reference we can guarantee that it won't
1462 	 * increase while we have the map locked.  Otherwise the race is
1463 	 * harmless and we will end up with an extra shadow object that
1464 	 * will be collapsed later.
1465 	 */
1466 	if (source != NULL && source->ref_count == 1 &&
1467 	    (source->flags & OBJ_ANON) != 0)
1468 		return;
1469 
1470 	/*
1471 	 * Allocate a new object with the given length.
1472 	 */
1473 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1474 
1475 	/*
1476 	 * Store the offset into the source object, and fix up the offset into
1477 	 * the new object.
1478 	 */
1479 	result->backing_object_offset = *offset;
1480 
1481 	if (shared || source != NULL) {
1482 		VM_OBJECT_WLOCK(result);
1483 
1484 		/*
1485 		 * The new object shadows the source object, adding a
1486 		 * reference to it.  Our caller changes his reference
1487 		 * to point to the new object, removing a reference to
1488 		 * the source object.  Net result: no change of
1489 		 * reference count, unless the caller needs to add one
1490 		 * more reference due to forking a shared map entry.
1491 		 */
1492 		if (shared) {
1493 			vm_object_reference_locked(result);
1494 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1495 		}
1496 
1497 		/*
1498 		 * Try to optimize the result object's page color when
1499 		 * shadowing in order to maintain page coloring
1500 		 * consistency in the combined shadowed object.
1501 		 */
1502 		if (source != NULL) {
1503 			vm_object_backing_insert(result, source);
1504 			result->domain = source->domain;
1505 #if VM_NRESERVLEVEL > 0
1506 			vm_object_set_flag(result,
1507 			    (source->flags & OBJ_COLORED));
1508 			result->pg_color = (source->pg_color +
1509 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1510 			    1)) - 1);
1511 #endif
1512 		}
1513 		VM_OBJECT_WUNLOCK(result);
1514 	}
1515 
1516 	/*
1517 	 * Return the new things
1518 	 */
1519 	*offset = 0;
1520 	*object = result;
1521 }
1522 
1523 /*
1524  *	vm_object_split:
1525  *
1526  * Split the pages in a map entry into a new object.  This affords
1527  * easier removal of unused pages, and keeps object inheritance from
1528  * being a negative impact on memory usage.
1529  */
1530 void
1531 vm_object_split(vm_map_entry_t entry)
1532 {
1533 	struct pctrie_iter pages;
1534 	vm_page_t m;
1535 	vm_object_t orig_object, new_object, backing_object;
1536 	vm_pindex_t offidxstart;
1537 	vm_size_t size;
1538 
1539 	orig_object = entry->object.vm_object;
1540 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1541 	    ("vm_object_split:  Splitting object with multiple mappings."));
1542 	if ((orig_object->flags & OBJ_ANON) == 0)
1543 		return;
1544 	if (orig_object->ref_count <= 1)
1545 		return;
1546 	VM_OBJECT_WUNLOCK(orig_object);
1547 
1548 	offidxstart = OFF_TO_IDX(entry->offset);
1549 	size = atop(entry->end - entry->start);
1550 
1551 	new_object = vm_object_allocate_anon(size, orig_object,
1552 	    orig_object->cred, ptoa(size));
1553 
1554 	/*
1555 	 * We must wait for the orig_object to complete any in-progress
1556 	 * collapse so that the swap blocks are stable below.  The
1557 	 * additional reference on backing_object by new object will
1558 	 * prevent further collapse operations until split completes.
1559 	 */
1560 	VM_OBJECT_WLOCK(orig_object);
1561 	vm_object_collapse_wait(orig_object);
1562 
1563 	/*
1564 	 * At this point, the new object is still private, so the order in
1565 	 * which the original and new objects are locked does not matter.
1566 	 */
1567 	VM_OBJECT_WLOCK(new_object);
1568 	new_object->domain = orig_object->domain;
1569 	backing_object = orig_object->backing_object;
1570 	if (backing_object != NULL) {
1571 		vm_object_backing_insert_ref(new_object, backing_object);
1572 		new_object->backing_object_offset =
1573 		    orig_object->backing_object_offset + entry->offset;
1574 	}
1575 	if (orig_object->cred != NULL) {
1576 		crhold(orig_object->cred);
1577 		KASSERT(orig_object->charge >= ptoa(size),
1578 		    ("orig_object->charge < 0"));
1579 		orig_object->charge -= ptoa(size);
1580 	}
1581 
1582 	/*
1583 	 * Mark the split operation so that swap_pager_getpages() knows
1584 	 * that the object is in transition.
1585 	 */
1586 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1587 	vm_page_iter_limit_init(&pages, orig_object, offidxstart + size);
1588 retry:
1589 	KASSERT(pctrie_iter_is_reset(&pages),
1590 	    ("%s: pctrie_iter not reset for retry", __func__));
1591 	for (m = vm_radix_iter_lookup_ge(&pages, offidxstart); m != NULL;
1592 	    m = vm_radix_iter_step(&pages)) {
1593 		/*
1594 		 * We must wait for pending I/O to complete before we can
1595 		 * rename the page.
1596 		 *
1597 		 * We do not have to VM_PROT_NONE the page as mappings should
1598 		 * not be changed by this operation.
1599 		 */
1600 		if (vm_page_tryxbusy(m) == 0) {
1601 			VM_OBJECT_WUNLOCK(new_object);
1602 			if (vm_page_busy_sleep(m, "spltwt", 0))
1603 				VM_OBJECT_WLOCK(orig_object);
1604 			pctrie_iter_reset(&pages);
1605 			VM_OBJECT_WLOCK(new_object);
1606 			goto retry;
1607 		}
1608 
1609 		/*
1610 		 * If the page was left invalid, it was likely placed there by
1611 		 * an incomplete fault.  Just remove and ignore.
1612 		 *
1613 		 * One other possibility is that the map entry is wired, in
1614 		 * which case we must hang on to the page to avoid leaking it,
1615 		 * as the map entry owns the wiring.  This case can arise if the
1616 		 * backing object is truncated by the pager.
1617 		 */
1618 		if (vm_page_none_valid(m) && entry->wired_count == 0) {
1619 			if (vm_page_iter_remove(&pages, m))
1620 				vm_page_free(m);
1621 			continue;
1622 		}
1623 
1624 		/* vm_page_iter_rename() will dirty the page if it is valid. */
1625 		if (!vm_page_iter_rename(&pages, m, new_object, m->pindex -
1626 		    offidxstart)) {
1627 			vm_page_xunbusy(m);
1628 			VM_OBJECT_WUNLOCK(new_object);
1629 			VM_OBJECT_WUNLOCK(orig_object);
1630 			vm_radix_wait();
1631 			pctrie_iter_reset(&pages);
1632 			VM_OBJECT_WLOCK(orig_object);
1633 			VM_OBJECT_WLOCK(new_object);
1634 			goto retry;
1635 		}
1636 
1637 #if VM_NRESERVLEVEL > 0
1638 		/*
1639 		 * If some of the reservation's allocated pages remain with
1640 		 * the original object, then transferring the reservation to
1641 		 * the new object is neither particularly beneficial nor
1642 		 * particularly harmful as compared to leaving the reservation
1643 		 * with the original object.  If, however, all of the
1644 		 * reservation's allocated pages are transferred to the new
1645 		 * object, then transferring the reservation is typically
1646 		 * beneficial.  Determining which of these two cases applies
1647 		 * would be more costly than unconditionally renaming the
1648 		 * reservation.
1649 		 */
1650 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1651 #endif
1652 	}
1653 
1654 	/*
1655 	 * swap_pager_copy() can sleep, in which case the orig_object's
1656 	 * and new_object's locks are released and reacquired.
1657 	 */
1658 	swap_pager_copy(orig_object, new_object, offidxstart, 0);
1659 	vm_page_iter_init(&pages, new_object);
1660 	VM_RADIX_FOREACH(m, &pages)
1661 		vm_page_xunbusy(m);
1662 
1663 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1664 	VM_OBJECT_WUNLOCK(orig_object);
1665 	VM_OBJECT_WUNLOCK(new_object);
1666 	entry->object.vm_object = new_object;
1667 	entry->offset = 0LL;
1668 	vm_object_deallocate(orig_object);
1669 	VM_OBJECT_WLOCK(new_object);
1670 }
1671 
1672 static vm_page_t
1673 vm_object_collapse_scan_wait(struct pctrie_iter *pages, vm_object_t object,
1674     vm_page_t p)
1675 {
1676 	vm_object_t backing_object;
1677 
1678 	VM_OBJECT_ASSERT_WLOCKED(object);
1679 	backing_object = object->backing_object;
1680 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1681 
1682 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1683 	    ("invalid ownership %p %p %p", p, object, backing_object));
1684 	/* The page is only NULL when rename fails. */
1685 	if (p == NULL) {
1686 		VM_OBJECT_WUNLOCK(object);
1687 		VM_OBJECT_WUNLOCK(backing_object);
1688 		vm_radix_wait();
1689 		VM_OBJECT_WLOCK(object);
1690 	} else if (p->object == object) {
1691 		VM_OBJECT_WUNLOCK(backing_object);
1692 		if (vm_page_busy_sleep(p, "vmocol", 0))
1693 			VM_OBJECT_WLOCK(object);
1694 	} else {
1695 		VM_OBJECT_WUNLOCK(object);
1696 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1697 			VM_OBJECT_WUNLOCK(backing_object);
1698 		VM_OBJECT_WLOCK(object);
1699 	}
1700 	VM_OBJECT_WLOCK(backing_object);
1701 	vm_page_iter_init(pages, backing_object);
1702 	return (vm_radix_iter_lookup_ge(pages, 0));
1703 }
1704 
1705 static void
1706 vm_object_collapse_scan(vm_object_t object)
1707 {
1708 	struct pctrie_iter pages;
1709 	vm_object_t backing_object;
1710 	vm_page_t next, p, pp;
1711 	vm_pindex_t backing_offset_index, new_pindex;
1712 
1713 	VM_OBJECT_ASSERT_WLOCKED(object);
1714 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1715 
1716 	backing_object = object->backing_object;
1717 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1718 
1719 	/*
1720 	 * Our scan
1721 	 */
1722 	vm_page_iter_init(&pages, backing_object);
1723 	for (p = vm_radix_iter_lookup_ge(&pages, 0); p != NULL; p = next) {
1724 		/*
1725 		 * Check for busy page
1726 		 */
1727 		if (vm_page_tryxbusy(p) == 0) {
1728 			next = vm_object_collapse_scan_wait(&pages, object, p);
1729 			continue;
1730 		}
1731 
1732 		KASSERT(object->backing_object == backing_object,
1733 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1734 		    object->backing_object, backing_object));
1735 		KASSERT(p->object == backing_object,
1736 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1737 		    p->object, backing_object));
1738 
1739 		if (p->pindex < backing_offset_index || object->size <=
1740 		    (new_pindex = p->pindex - backing_offset_index)) {
1741 			vm_pager_freespace(backing_object, p->pindex, 1);
1742 
1743 			KASSERT(!pmap_page_is_mapped(p),
1744 			    ("freeing mapped page %p", p));
1745 			if (vm_page_iter_remove(&pages, p))
1746 				vm_page_free(p);
1747 			next = vm_radix_iter_step(&pages);
1748 			continue;
1749 		}
1750 
1751 		if (!vm_page_all_valid(p)) {
1752 			KASSERT(!pmap_page_is_mapped(p),
1753 			    ("freeing mapped page %p", p));
1754 			if (vm_page_iter_remove(&pages, p))
1755 				vm_page_free(p);
1756 			next = vm_radix_iter_step(&pages);
1757 			continue;
1758 		}
1759 
1760 		pp = vm_page_lookup(object, new_pindex);
1761 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1762 			vm_page_xunbusy(p);
1763 			/*
1764 			 * The page in the parent is busy and possibly not
1765 			 * (yet) valid.  Until its state is finalized by the
1766 			 * busy bit owner, we can't tell whether it shadows the
1767 			 * original page.
1768 			 */
1769 			next = vm_object_collapse_scan_wait(&pages, object, pp);
1770 			continue;
1771 		}
1772 
1773 		if (pp != NULL && vm_page_none_valid(pp)) {
1774 			/*
1775 			 * The page was invalid in the parent.  Likely placed
1776 			 * there by an incomplete fault.  Just remove and
1777 			 * ignore.  p can replace it.
1778 			 */
1779 			if (vm_page_remove(pp))
1780 				vm_page_free(pp);
1781 			pp = NULL;
1782 		}
1783 
1784 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1785 			NULL)) {
1786 			/*
1787 			 * The page already exists in the parent OR swap exists
1788 			 * for this location in the parent.  Leave the parent's
1789 			 * page alone.  Destroy the original page from the
1790 			 * backing object.
1791 			 */
1792 			vm_pager_freespace(backing_object, p->pindex, 1);
1793 			KASSERT(!pmap_page_is_mapped(p),
1794 			    ("freeing mapped page %p", p));
1795 			if (pp != NULL)
1796 				vm_page_xunbusy(pp);
1797 			if (vm_page_iter_remove(&pages, p))
1798 				vm_page_free(p);
1799 			next = vm_radix_iter_step(&pages);
1800 			continue;
1801 		}
1802 
1803 		/*
1804 		 * Page does not exist in parent, rename the page from the
1805 		 * backing object to the main object.
1806 		 *
1807 		 * If the page was mapped to a process, it can remain mapped
1808 		 * through the rename.  vm_page_iter_rename() will dirty the
1809 		 * page.
1810 		 */
1811 		if (!vm_page_iter_rename(&pages, p, object, new_pindex)) {
1812 			vm_page_xunbusy(p);
1813 			next = vm_object_collapse_scan_wait(&pages, object,
1814 			    NULL);
1815 			continue;
1816 		}
1817 
1818 		/* Use the old pindex to free the right page. */
1819 		vm_pager_freespace(backing_object, new_pindex +
1820 		    backing_offset_index, 1);
1821 
1822 #if VM_NRESERVLEVEL > 0
1823 		/*
1824 		 * Rename the reservation.
1825 		 */
1826 		vm_reserv_rename(p, object, backing_object,
1827 		    backing_offset_index);
1828 #endif
1829 		vm_page_xunbusy(p);
1830 		next = vm_radix_iter_step(&pages);
1831 	}
1832 	return;
1833 }
1834 
1835 /*
1836  *	vm_object_collapse:
1837  *
1838  *	Collapse an object with the object backing it.
1839  *	Pages in the backing object are moved into the
1840  *	parent, and the backing object is deallocated.
1841  */
1842 void
1843 vm_object_collapse(vm_object_t object)
1844 {
1845 	vm_object_t backing_object, new_backing_object;
1846 
1847 	VM_OBJECT_ASSERT_WLOCKED(object);
1848 
1849 	while (TRUE) {
1850 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1851 		    ("collapsing invalid object"));
1852 
1853 		/*
1854 		 * Wait for the backing_object to finish any pending
1855 		 * collapse so that the caller sees the shortest possible
1856 		 * shadow chain.
1857 		 */
1858 		backing_object = vm_object_backing_collapse_wait(object);
1859 		if (backing_object == NULL)
1860 			return;
1861 
1862 		KASSERT(object->ref_count > 0 &&
1863 		    object->ref_count > atomic_load_int(&object->shadow_count),
1864 		    ("collapse with invalid ref %d or shadow %d count.",
1865 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1866 		KASSERT((backing_object->flags &
1867 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1868 		    ("vm_object_collapse: Backing object already collapsing."));
1869 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1870 		    ("vm_object_collapse: object is already collapsing."));
1871 
1872 		/*
1873 		 * We know that we can either collapse the backing object if
1874 		 * the parent is the only reference to it, or (perhaps) have
1875 		 * the parent bypass the object if the parent happens to shadow
1876 		 * all the resident pages in the entire backing object.
1877 		 */
1878 		if (backing_object->ref_count == 1) {
1879 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1880 			    == 1,
1881 			    ("vm_object_collapse: shadow_count: %d",
1882 			    atomic_load_int(&backing_object->shadow_count)));
1883 			vm_object_pip_add(object, 1);
1884 			vm_object_set_flag(object, OBJ_COLLAPSING);
1885 			vm_object_pip_add(backing_object, 1);
1886 			vm_object_set_flag(backing_object, OBJ_DEAD);
1887 
1888 			/*
1889 			 * If there is exactly one reference to the backing
1890 			 * object, we can collapse it into the parent.
1891 			 */
1892 			vm_object_collapse_scan(object);
1893 
1894 			/*
1895 			 * Move the pager from backing_object to object.
1896 			 *
1897 			 * swap_pager_copy() can sleep, in which case the
1898 			 * backing_object's and object's locks are released and
1899 			 * reacquired.
1900 			 */
1901 			swap_pager_copy(backing_object, object,
1902 			    OFF_TO_IDX(object->backing_object_offset), TRUE);
1903 
1904 			/*
1905 			 * Object now shadows whatever backing_object did.
1906 			 */
1907 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1908 			vm_object_backing_transfer(object, backing_object);
1909 			object->backing_object_offset +=
1910 			    backing_object->backing_object_offset;
1911 			VM_OBJECT_WUNLOCK(object);
1912 			vm_object_pip_wakeup(object);
1913 
1914 			/*
1915 			 * Discard backing_object.
1916 			 *
1917 			 * Since the backing object has no pages, no pager left,
1918 			 * and no object references within it, all that is
1919 			 * necessary is to dispose of it.
1920 			 */
1921 			KASSERT(backing_object->ref_count == 1, (
1922 "backing_object %p was somehow re-referenced during collapse!",
1923 			    backing_object));
1924 			vm_object_pip_wakeup(backing_object);
1925 			(void)refcount_release(&backing_object->ref_count);
1926 			umtx_shm_object_terminated(backing_object);
1927 			vm_object_terminate(backing_object);
1928 			counter_u64_add(object_collapses, 1);
1929 			VM_OBJECT_WLOCK(object);
1930 		} else {
1931 			/*
1932 			 * If we do not entirely shadow the backing object,
1933 			 * there is nothing we can do so we give up.
1934 			 *
1935 			 * The object lock and backing_object lock must not
1936 			 * be dropped during this sequence.
1937 			 */
1938 			if (!swap_pager_scan_all_shadowed(object)) {
1939 				VM_OBJECT_WUNLOCK(backing_object);
1940 				break;
1941 			}
1942 
1943 			/*
1944 			 * Make the parent shadow the next object in the
1945 			 * chain.  Deallocating backing_object will not remove
1946 			 * it, since its reference count is at least 2.
1947 			 */
1948 			vm_object_backing_remove_locked(object);
1949 			new_backing_object = backing_object->backing_object;
1950 			if (new_backing_object != NULL) {
1951 				vm_object_backing_insert_ref(object,
1952 				    new_backing_object);
1953 				object->backing_object_offset +=
1954 				    backing_object->backing_object_offset;
1955 			}
1956 
1957 			/*
1958 			 * Drop the reference count on backing_object. Since
1959 			 * its ref_count was at least 2, it will not vanish.
1960 			 */
1961 			(void)refcount_release(&backing_object->ref_count);
1962 			KASSERT(backing_object->ref_count >= 1, (
1963 "backing_object %p was somehow dereferenced during collapse!",
1964 			    backing_object));
1965 			VM_OBJECT_WUNLOCK(backing_object);
1966 			counter_u64_add(object_bypasses, 1);
1967 		}
1968 
1969 		/*
1970 		 * Try again with this object's new backing object.
1971 		 */
1972 	}
1973 }
1974 
1975 /*
1976  *	vm_object_page_remove:
1977  *
1978  *	For the given object, either frees or invalidates each of the
1979  *	specified pages.  In general, a page is freed.  However, if a page is
1980  *	wired for any reason other than the existence of a managed, wired
1981  *	mapping, then it may be invalidated but not removed from the object.
1982  *	Pages are specified by the given range ["start", "end") and the option
1983  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1984  *	extends from "start" to the end of the object.  If the option
1985  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1986  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
1987  *	specified, then the pages within the specified range must have no
1988  *	mappings.  Otherwise, if this option is not specified, any mappings to
1989  *	the specified pages are removed before the pages are freed or
1990  *	invalidated.
1991  *
1992  *	In general, this operation should only be performed on objects that
1993  *	contain managed pages.  There are, however, two exceptions.  First, it
1994  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
1995  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1996  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1997  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
1998  *
1999  *	The object must be locked.
2000  */
2001 void
2002 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2003     int options)
2004 {
2005 	struct pctrie_iter pages;
2006 	vm_page_t p;
2007 
2008 	VM_OBJECT_ASSERT_WLOCKED(object);
2009 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2010 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2011 	    ("vm_object_page_remove: illegal options for object %p", object));
2012 	if (object->resident_page_count == 0)
2013 		return;
2014 	vm_object_pip_add(object, 1);
2015 	vm_page_iter_limit_init(&pages, object, end);
2016 again:
2017 	KASSERT(pctrie_iter_is_reset(&pages),
2018 	    ("%s: pctrie_iter not reset for retry", __func__));
2019 	for (p = vm_radix_iter_lookup_ge(&pages, start); p != NULL;
2020 	     p = vm_radix_iter_step(&pages)) {
2021 		/*
2022 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2023 		 * the busy lock, as some consumers rely on this to avoid
2024 		 * deadlocks.
2025 		 *
2026 		 * A thread may concurrently transition the page from invalid to
2027 		 * valid using only the busy lock, so the result of this check
2028 		 * is immediately stale.  It is up to consumers to handle this,
2029 		 * for instance by ensuring that all invalid->valid transitions
2030 		 * happen with a mutex held, as may be possible for a
2031 		 * filesystem.
2032 		 */
2033 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2034 			continue;
2035 
2036 		/*
2037 		 * If the page is wired for any reason besides the existence
2038 		 * of managed, wired mappings, then it cannot be freed.  For
2039 		 * example, fictitious pages, which represent device memory,
2040 		 * are inherently wired and cannot be freed.  They can,
2041 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2042 		 * not specified.
2043 		 */
2044 		if (vm_page_tryxbusy(p) == 0) {
2045 			if (vm_page_busy_sleep(p, "vmopar", 0))
2046 				VM_OBJECT_WLOCK(object);
2047 			pctrie_iter_reset(&pages);
2048 			goto again;
2049 		}
2050 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2051 			vm_page_xunbusy(p);
2052 			continue;
2053 		}
2054 		if (vm_page_wired(p)) {
2055 wired:
2056 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2057 			    object->ref_count != 0)
2058 				pmap_remove_all(p);
2059 			if ((options & OBJPR_CLEANONLY) == 0) {
2060 				vm_page_invalid(p);
2061 				vm_page_undirty(p);
2062 			}
2063 			vm_page_xunbusy(p);
2064 			continue;
2065 		}
2066 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2067 		    ("vm_object_page_remove: page %p is fictitious", p));
2068 		if ((options & OBJPR_CLEANONLY) != 0 &&
2069 		    !vm_page_none_valid(p)) {
2070 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2071 			    object->ref_count != 0 &&
2072 			    !vm_page_try_remove_write(p))
2073 				goto wired;
2074 			if (p->dirty != 0) {
2075 				vm_page_xunbusy(p);
2076 				continue;
2077 			}
2078 		}
2079 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2080 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2081 			goto wired;
2082 		vm_page_iter_free(&pages, p);
2083 	}
2084 	vm_object_pip_wakeup(object);
2085 
2086 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2087 	    start);
2088 }
2089 
2090 /*
2091  *	vm_object_page_noreuse:
2092  *
2093  *	For the given object, attempt to move the specified pages to
2094  *	the head of the inactive queue.  This bypasses regular LRU
2095  *	operation and allows the pages to be reused quickly under memory
2096  *	pressure.  If a page is wired for any reason, then it will not
2097  *	be queued.  Pages are specified by the range ["start", "end").
2098  *	As a special case, if "end" is zero, then the range extends from
2099  *	"start" to the end of the object.
2100  *
2101  *	This operation should only be performed on objects that
2102  *	contain non-fictitious, managed pages.
2103  *
2104  *	The object must be locked.
2105  */
2106 void
2107 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2108 {
2109 	struct pctrie_iter pages;
2110 	vm_page_t p;
2111 
2112 	VM_OBJECT_ASSERT_LOCKED(object);
2113 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2114 	    ("vm_object_page_noreuse: illegal object %p", object));
2115 	if (object->resident_page_count == 0)
2116 		return;
2117 
2118 	vm_page_iter_limit_init(&pages, object, end);
2119 	VM_RADIX_FOREACH_FROM(p, &pages, start)
2120 		vm_page_deactivate_noreuse(p);
2121 }
2122 
2123 /*
2124  *	Populate the specified range of the object with valid pages.  Returns
2125  *	TRUE if the range is successfully populated and FALSE otherwise.
2126  *
2127  *	Note: This function should be optimized to pass a larger array of
2128  *	pages to vm_pager_get_pages() before it is applied to a non-
2129  *	OBJT_DEVICE object.
2130  *
2131  *	The object must be locked.
2132  */
2133 boolean_t
2134 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2135 {
2136 	struct pctrie_iter pages;
2137 	vm_page_t m;
2138 	vm_pindex_t pindex;
2139 	int rv;
2140 
2141 	vm_page_iter_init(&pages, object);
2142 	VM_OBJECT_ASSERT_WLOCKED(object);
2143 	for (pindex = start; pindex < end; pindex++) {
2144 		rv = vm_page_grab_valid_iter(&m, object, &pages, pindex,
2145 		    VM_ALLOC_NORMAL);
2146 		if (rv != VM_PAGER_OK)
2147 			break;
2148 
2149 		/*
2150 		 * Keep "m" busy because a subsequent iteration may unlock
2151 		 * the object.
2152 		 */
2153 	}
2154 	if (pindex > start) {
2155 		pages.limit = pindex;
2156 		VM_RADIX_FORALL_FROM(m, &pages, start)
2157 			vm_page_xunbusy(m);
2158 	}
2159 	return (pindex == end);
2160 }
2161 
2162 /*
2163  *	Routine:	vm_object_coalesce
2164  *	Function:	Coalesces two objects backing up adjoining
2165  *			regions of memory into a single object.
2166  *
2167  *	returns TRUE if objects were combined.
2168  *
2169  *	NOTE:	Only works at the moment if the second object is NULL -
2170  *		if it's not, which object do we lock first?
2171  *
2172  *	Parameters:
2173  *		prev_object	First object to coalesce
2174  *		prev_offset	Offset into prev_object
2175  *		prev_size	Size of reference to prev_object
2176  *		next_size	Size of reference to the second object
2177  *		reserved	Indicator that extension region has
2178  *				swap accounted for
2179  *
2180  *	Conditions:
2181  *	The object must *not* be locked.
2182  */
2183 boolean_t
2184 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2185     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2186 {
2187 	vm_pindex_t next_pindex;
2188 
2189 	if (prev_object == NULL)
2190 		return (TRUE);
2191 	if ((prev_object->flags & OBJ_ANON) == 0)
2192 		return (FALSE);
2193 
2194 	VM_OBJECT_WLOCK(prev_object);
2195 	/*
2196 	 * Try to collapse the object first.
2197 	 */
2198 	vm_object_collapse(prev_object);
2199 
2200 	/*
2201 	 * Can't coalesce if: . more than one reference . paged out . shadows
2202 	 * another object . has a copy elsewhere (any of which mean that the
2203 	 * pages not mapped to prev_entry may be in use anyway)
2204 	 */
2205 	if (prev_object->backing_object != NULL) {
2206 		VM_OBJECT_WUNLOCK(prev_object);
2207 		return (FALSE);
2208 	}
2209 
2210 	prev_size >>= PAGE_SHIFT;
2211 	next_size >>= PAGE_SHIFT;
2212 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2213 
2214 	if (prev_object->ref_count > 1 &&
2215 	    prev_object->size != next_pindex &&
2216 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2217 		VM_OBJECT_WUNLOCK(prev_object);
2218 		return (FALSE);
2219 	}
2220 
2221 	/*
2222 	 * Account for the charge.
2223 	 */
2224 	if (prev_object->cred != NULL) {
2225 		/*
2226 		 * If prev_object was charged, then this mapping,
2227 		 * although not charged now, may become writable
2228 		 * later. Non-NULL cred in the object would prevent
2229 		 * swap reservation during enabling of the write
2230 		 * access, so reserve swap now. Failed reservation
2231 		 * cause allocation of the separate object for the map
2232 		 * entry, and swap reservation for this entry is
2233 		 * managed in appropriate time.
2234 		 */
2235 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2236 		    prev_object->cred)) {
2237 			VM_OBJECT_WUNLOCK(prev_object);
2238 			return (FALSE);
2239 		}
2240 		prev_object->charge += ptoa(next_size);
2241 	}
2242 
2243 	/*
2244 	 * Remove any pages that may still be in the object from a previous
2245 	 * deallocation.
2246 	 */
2247 	if (next_pindex < prev_object->size) {
2248 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2249 		    next_size, 0);
2250 #if 0
2251 		if (prev_object->cred != NULL) {
2252 			KASSERT(prev_object->charge >=
2253 			    ptoa(prev_object->size - next_pindex),
2254 			    ("object %p overcharged 1 %jx %jx", prev_object,
2255 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2256 			prev_object->charge -= ptoa(prev_object->size -
2257 			    next_pindex);
2258 		}
2259 #endif
2260 	}
2261 
2262 	/*
2263 	 * Extend the object if necessary.
2264 	 */
2265 	if (next_pindex + next_size > prev_object->size)
2266 		prev_object->size = next_pindex + next_size;
2267 
2268 	VM_OBJECT_WUNLOCK(prev_object);
2269 	return (TRUE);
2270 }
2271 
2272 /*
2273  * Fill in the m_dst array with up to *rbehind optional pages before m_src[0]
2274  * and up to *rahead optional pages after m_src[count - 1].  In both cases, stop
2275  * the filling-in short on encountering a cached page, an object boundary limit,
2276  * or an allocation error.  Update *rbehind and *rahead to indicate the number
2277  * of pages allocated.  Copy elements of m_src into array elements from
2278  * m_dst[*rbehind] to m_dst[*rbehind + count -1].
2279  */
2280 void
2281 vm_object_prepare_buf_pages(vm_object_t object, vm_page_t *ma_dst, int count,
2282     int *rbehind, int *rahead, vm_page_t *ma_src)
2283 {
2284 	struct pctrie_iter pages;
2285 	vm_pindex_t pindex;
2286 	vm_page_t m, mpred, msucc;
2287 
2288 	vm_page_iter_init(&pages, object);
2289 	VM_OBJECT_ASSERT_LOCKED(object);
2290 	if (*rbehind != 0) {
2291 		m = ma_src[0];
2292 		pindex = m->pindex;
2293 		mpred = vm_radix_iter_lookup_lt(&pages, pindex);
2294 		*rbehind = MIN(*rbehind,
2295 		    pindex - (mpred != NULL ? mpred->pindex + 1 : 0));
2296 		/* Stepping backward from pindex, mpred doesn't change. */
2297 		for (int i = 0; i < *rbehind; i++) {
2298 			m = vm_page_alloc_after(object, &pages, pindex - i - 1,
2299 			    VM_ALLOC_NORMAL, mpred);
2300 			if (m == NULL) {
2301 				/* Shift the array. */
2302 				for (int j = 0; j < i; j++)
2303 					ma_dst[j] = ma_dst[j + *rbehind - i];
2304 				*rbehind = i;
2305 				*rahead = 0;
2306 				break;
2307 			}
2308 			ma_dst[*rbehind - i - 1] = m;
2309 		}
2310 	}
2311 	for (int i = 0; i < count; i++)
2312 		ma_dst[*rbehind + i] = ma_src[i];
2313 	if (*rahead != 0) {
2314 		m = ma_src[count - 1];
2315 		pindex = m->pindex + 1;
2316 		msucc = vm_radix_iter_lookup_ge(&pages, pindex);
2317 		*rahead = MIN(*rahead,
2318 		    (msucc != NULL ? msucc->pindex : object->size) - pindex);
2319 		mpred = m;
2320 		for (int i = 0; i < *rahead; i++) {
2321 			m = vm_page_alloc_after(object, &pages, pindex + i,
2322 			    VM_ALLOC_NORMAL, mpred);
2323 			if (m == NULL) {
2324 				*rahead = i;
2325 				break;
2326 			}
2327 			ma_dst[*rbehind + count + i] = mpred = m;
2328 		}
2329 	}
2330 }
2331 
2332 void
2333 vm_object_set_writeable_dirty_(vm_object_t object)
2334 {
2335 	atomic_add_int(&object->generation, 1);
2336 }
2337 
2338 bool
2339 vm_object_mightbedirty_(vm_object_t object)
2340 {
2341 	return (object->generation != object->cleangeneration);
2342 }
2343 
2344 /*
2345  *	vm_object_unwire:
2346  *
2347  *	For each page offset within the specified range of the given object,
2348  *	find the highest-level page in the shadow chain and unwire it.  A page
2349  *	must exist at every page offset, and the highest-level page must be
2350  *	wired.
2351  */
2352 void
2353 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2354     uint8_t queue)
2355 {
2356 	struct pctrie_iter pages;
2357 	vm_object_t tobject, t1object;
2358 	vm_page_t m, tm;
2359 	vm_pindex_t end_pindex, pindex, tpindex;
2360 	int depth, locked_depth;
2361 
2362 	KASSERT((offset & PAGE_MASK) == 0,
2363 	    ("vm_object_unwire: offset is not page aligned"));
2364 	KASSERT((length & PAGE_MASK) == 0,
2365 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2366 	/* The wired count of a fictitious page never changes. */
2367 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2368 		return;
2369 	pindex = OFF_TO_IDX(offset);
2370 	end_pindex = pindex + atop(length);
2371 	vm_page_iter_init(&pages, object);
2372 again:
2373 	locked_depth = 1;
2374 	VM_OBJECT_RLOCK(object);
2375 	m = vm_radix_iter_lookup_ge(&pages, pindex);
2376 	while (pindex < end_pindex) {
2377 		if (m == NULL || pindex < m->pindex) {
2378 			/*
2379 			 * The first object in the shadow chain doesn't
2380 			 * contain a page at the current index.  Therefore,
2381 			 * the page must exist in a backing object.
2382 			 */
2383 			tobject = object;
2384 			tpindex = pindex;
2385 			depth = 0;
2386 			do {
2387 				tpindex +=
2388 				    OFF_TO_IDX(tobject->backing_object_offset);
2389 				tobject = tobject->backing_object;
2390 				KASSERT(tobject != NULL,
2391 				    ("vm_object_unwire: missing page"));
2392 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2393 					goto next_page;
2394 				depth++;
2395 				if (depth == locked_depth) {
2396 					locked_depth++;
2397 					VM_OBJECT_RLOCK(tobject);
2398 				}
2399 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2400 			    NULL);
2401 		} else {
2402 			tm = m;
2403 			m = vm_radix_iter_step(&pages);
2404 		}
2405 		if (vm_page_trysbusy(tm) == 0) {
2406 			for (tobject = object; locked_depth >= 1;
2407 			    locked_depth--) {
2408 				t1object = tobject->backing_object;
2409 				if (tm->object != tobject)
2410 					VM_OBJECT_RUNLOCK(tobject);
2411 				tobject = t1object;
2412 			}
2413 			tobject = tm->object;
2414 			if (!vm_page_busy_sleep(tm, "unwbo",
2415 			    VM_ALLOC_IGN_SBUSY))
2416 				VM_OBJECT_RUNLOCK(tobject);
2417 			pctrie_iter_reset(&pages);
2418 			goto again;
2419 		}
2420 		vm_page_unwire(tm, queue);
2421 		vm_page_sunbusy(tm);
2422 next_page:
2423 		pindex++;
2424 	}
2425 	/* Release the accumulated object locks. */
2426 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2427 		t1object = tobject->backing_object;
2428 		VM_OBJECT_RUNLOCK(tobject);
2429 		tobject = t1object;
2430 	}
2431 }
2432 
2433 /*
2434  * Return the vnode for the given object, or NULL if none exists.
2435  * For tmpfs objects, the function may return NULL if there is
2436  * no vnode allocated at the time of the call.
2437  */
2438 struct vnode *
2439 vm_object_vnode(vm_object_t object)
2440 {
2441 	struct vnode *vp;
2442 
2443 	VM_OBJECT_ASSERT_LOCKED(object);
2444 	vm_pager_getvp(object, &vp, NULL);
2445 	return (vp);
2446 }
2447 
2448 /*
2449  * Busy the vm object.  This prevents new pages belonging to the object from
2450  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2451  * for checking page state before proceeding.
2452  */
2453 void
2454 vm_object_busy(vm_object_t obj)
2455 {
2456 
2457 	VM_OBJECT_ASSERT_LOCKED(obj);
2458 
2459 	blockcount_acquire(&obj->busy, 1);
2460 	/* The fence is required to order loads of page busy. */
2461 	atomic_thread_fence_acq_rel();
2462 }
2463 
2464 void
2465 vm_object_unbusy(vm_object_t obj)
2466 {
2467 
2468 	blockcount_release(&obj->busy, 1);
2469 }
2470 
2471 void
2472 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2473 {
2474 
2475 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2476 
2477 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2478 }
2479 
2480 /*
2481  * This function aims to determine if the object is mapped,
2482  * specifically, if it is referenced by a vm_map_entry.  Because
2483  * objects occasionally acquire transient references that do not
2484  * represent a mapping, the method used here is inexact.  However, it
2485  * has very low overhead and is good enough for the advisory
2486  * vm.vmtotal sysctl.
2487  */
2488 bool
2489 vm_object_is_active(vm_object_t obj)
2490 {
2491 
2492 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2493 }
2494 
2495 static int
2496 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2497 {
2498 	struct pctrie_iter pages;
2499 	struct kinfo_vmobject *kvo;
2500 	char *fullpath, *freepath;
2501 	struct vnode *vp;
2502 	struct vattr va;
2503 	vm_object_t obj;
2504 	vm_page_t m;
2505 	u_long sp;
2506 	int count, error;
2507 	key_t key;
2508 	unsigned short seq;
2509 	bool want_path;
2510 
2511 	if (req->oldptr == NULL) {
2512 		/*
2513 		 * If an old buffer has not been provided, generate an
2514 		 * estimate of the space needed for a subsequent call.
2515 		 */
2516 		mtx_lock(&vm_object_list_mtx);
2517 		count = 0;
2518 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2519 			if (obj->type == OBJT_DEAD)
2520 				continue;
2521 			count++;
2522 		}
2523 		mtx_unlock(&vm_object_list_mtx);
2524 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2525 		    count * 11 / 10));
2526 	}
2527 
2528 	want_path = !(swap_only || jailed(curthread->td_ucred));
2529 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2530 	error = 0;
2531 
2532 	/*
2533 	 * VM objects are type stable and are never removed from the
2534 	 * list once added.  This allows us to safely read obj->object_list
2535 	 * after reacquiring the VM object lock.
2536 	 */
2537 	mtx_lock(&vm_object_list_mtx);
2538 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2539 		if (obj->type == OBJT_DEAD ||
2540 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2541 			continue;
2542 		VM_OBJECT_RLOCK(obj);
2543 		if (obj->type == OBJT_DEAD ||
2544 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2545 			VM_OBJECT_RUNLOCK(obj);
2546 			continue;
2547 		}
2548 		mtx_unlock(&vm_object_list_mtx);
2549 		kvo->kvo_size = ptoa(obj->size);
2550 		kvo->kvo_resident = obj->resident_page_count;
2551 		kvo->kvo_ref_count = obj->ref_count;
2552 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2553 		kvo->kvo_memattr = obj->memattr;
2554 		kvo->kvo_active = 0;
2555 		kvo->kvo_inactive = 0;
2556 		kvo->kvo_flags = 0;
2557 		if (!swap_only) {
2558 			vm_page_iter_init(&pages, obj);
2559 			VM_RADIX_FOREACH(m, &pages) {
2560 				/*
2561 				 * A page may belong to the object but be
2562 				 * dequeued and set to PQ_NONE while the
2563 				 * object lock is not held.  This makes the
2564 				 * reads of m->queue below racy, and we do not
2565 				 * count pages set to PQ_NONE.  However, this
2566 				 * sysctl is only meant to give an
2567 				 * approximation of the system anyway.
2568 				 */
2569 				if (vm_page_active(m))
2570 					kvo->kvo_active++;
2571 				else if (vm_page_inactive(m))
2572 					kvo->kvo_inactive++;
2573 				else if (vm_page_in_laundry(m))
2574 					kvo->kvo_laundry++;
2575 			}
2576 		}
2577 
2578 		kvo->kvo_vn_fileid = 0;
2579 		kvo->kvo_vn_fsid = 0;
2580 		kvo->kvo_vn_fsid_freebsd11 = 0;
2581 		freepath = NULL;
2582 		fullpath = "";
2583 		vp = NULL;
2584 		kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp :
2585 		    NULL);
2586 		if (vp != NULL) {
2587 			vref(vp);
2588 		} else if ((obj->flags & OBJ_ANON) != 0) {
2589 			MPASS(kvo->kvo_type == KVME_TYPE_SWAP);
2590 			kvo->kvo_me = (uintptr_t)obj;
2591 			/* tmpfs objs are reported as vnodes */
2592 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2593 			sp = swap_pager_swapped_pages(obj);
2594 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2595 		}
2596 		if (obj->type == OBJT_DEVICE || obj->type == OBJT_MGTDEVICE) {
2597 			cdev_pager_get_path(obj, kvo->kvo_path,
2598 			    sizeof(kvo->kvo_path));
2599 		}
2600 		VM_OBJECT_RUNLOCK(obj);
2601 		if ((obj->flags & OBJ_SYSVSHM) != 0) {
2602 			kvo->kvo_flags |= KVMO_FLAG_SYSVSHM;
2603 			shmobjinfo(obj, &key, &seq);
2604 			kvo->kvo_vn_fileid = key;
2605 			kvo->kvo_vn_fsid_freebsd11 = seq;
2606 		}
2607 		if ((obj->flags & OBJ_POSIXSHM) != 0) {
2608 			kvo->kvo_flags |= KVMO_FLAG_POSIXSHM;
2609 			shm_get_path(obj, kvo->kvo_path,
2610 			    sizeof(kvo->kvo_path));
2611 		}
2612 		if (vp != NULL) {
2613 			vn_fullpath(vp, &fullpath, &freepath);
2614 			vn_lock(vp, LK_SHARED | LK_RETRY);
2615 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2616 				kvo->kvo_vn_fileid = va.va_fileid;
2617 				kvo->kvo_vn_fsid = va.va_fsid;
2618 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2619 								/* truncate */
2620 			}
2621 			vput(vp);
2622 			strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2623 			free(freepath, M_TEMP);
2624 		}
2625 
2626 		/* Pack record size down */
2627 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2628 		    + strlen(kvo->kvo_path) + 1;
2629 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2630 		    sizeof(uint64_t));
2631 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2632 		maybe_yield();
2633 		mtx_lock(&vm_object_list_mtx);
2634 		if (error)
2635 			break;
2636 	}
2637 	mtx_unlock(&vm_object_list_mtx);
2638 	free(kvo, M_TEMP);
2639 	return (error);
2640 }
2641 
2642 static int
2643 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2644 {
2645 	return (vm_object_list_handler(req, false));
2646 }
2647 
2648 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2649     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2650     "List of VM objects");
2651 
2652 static int
2653 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2654 {
2655 	return (vm_object_list_handler(req, true));
2656 }
2657 
2658 /*
2659  * This sysctl returns list of the anonymous or swap objects. Intent
2660  * is to provide stripped optimized list useful to analyze swap use.
2661  * Since technically non-swap (default) objects participate in the
2662  * shadow chains, and are converted to swap type as needed by swap
2663  * pager, we must report them.
2664  */
2665 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2666     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2667     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2668     "List of swap VM objects");
2669 
2670 #include "opt_ddb.h"
2671 #ifdef DDB
2672 #include <sys/kernel.h>
2673 
2674 #include <sys/cons.h>
2675 
2676 #include <ddb/ddb.h>
2677 
2678 static int
2679 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2680 {
2681 	vm_map_t tmpm;
2682 	vm_map_entry_t tmpe;
2683 	vm_object_t obj;
2684 
2685 	if (map == 0)
2686 		return 0;
2687 
2688 	if (entry == 0) {
2689 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2690 			if (_vm_object_in_map(map, object, tmpe)) {
2691 				return 1;
2692 			}
2693 		}
2694 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2695 		tmpm = entry->object.sub_map;
2696 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2697 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2698 				return 1;
2699 			}
2700 		}
2701 	} else if ((obj = entry->object.vm_object) != NULL) {
2702 		for (; obj; obj = obj->backing_object)
2703 			if (obj == object) {
2704 				return 1;
2705 			}
2706 	}
2707 	return 0;
2708 }
2709 
2710 static int
2711 vm_object_in_map(vm_object_t object)
2712 {
2713 	struct proc *p;
2714 
2715 	/* sx_slock(&allproc_lock); */
2716 	FOREACH_PROC_IN_SYSTEM(p) {
2717 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2718 			continue;
2719 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2720 			/* sx_sunlock(&allproc_lock); */
2721 			return 1;
2722 		}
2723 	}
2724 	/* sx_sunlock(&allproc_lock); */
2725 	if (_vm_object_in_map(kernel_map, object, 0))
2726 		return 1;
2727 	return 0;
2728 }
2729 
2730 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE)
2731 {
2732 	vm_object_t object;
2733 
2734 	/*
2735 	 * make sure that internal objs are in a map somewhere
2736 	 * and none have zero ref counts.
2737 	 */
2738 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2739 		if ((object->flags & OBJ_ANON) != 0) {
2740 			if (object->ref_count == 0) {
2741 				db_printf(
2742 			"vmochk: internal obj has zero ref count: %lu\n",
2743 				    (u_long)object->size);
2744 			}
2745 			if (!vm_object_in_map(object)) {
2746 				db_printf(
2747 			"vmochk: internal obj is not in a map: "
2748 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2749 				    object->ref_count, (u_long)object->size,
2750 				    (u_long)object->size,
2751 				    (void *)object->backing_object);
2752 			}
2753 		}
2754 		if (db_pager_quit)
2755 			return;
2756 	}
2757 }
2758 
2759 /*
2760  *	vm_object_print:	[ debug ]
2761  */
2762 DB_SHOW_COMMAND(object, vm_object_print_static)
2763 {
2764 	struct pctrie_iter pages;
2765 	/* XXX convert args. */
2766 	vm_object_t object = (vm_object_t)addr;
2767 	boolean_t full = have_addr;
2768 
2769 	vm_page_t p;
2770 
2771 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2772 #define	count	was_count
2773 
2774 	int count;
2775 
2776 	if (object == NULL)
2777 		return;
2778 
2779 	db_iprintf("Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x",
2780 	    object, (int)object->type, (uintmax_t)object->size,
2781 	    object->resident_page_count, object->ref_count, object->flags);
2782 	db_iprintf(" ruid %d charge %jx\n",
2783 	    object->cred ? object->cred->cr_ruid : -1,
2784 	    (uintmax_t)object->charge);
2785 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2786 	    atomic_load_int(&object->shadow_count),
2787 	    object->backing_object ? object->backing_object->ref_count : 0,
2788 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2789 
2790 	if (!full)
2791 		return;
2792 
2793 	db_indent += 2;
2794 	count = 0;
2795 	vm_page_iter_init(&pages, object);
2796 	VM_RADIX_FOREACH(p, &pages) {
2797 		if (count == 0)
2798 			db_iprintf("memory:=");
2799 		else if (count == 6) {
2800 			db_printf("\n");
2801 			db_iprintf(" ...");
2802 			count = 0;
2803 		} else
2804 			db_printf(",");
2805 		count++;
2806 
2807 		db_printf("(off=0x%jx,page=0x%jx)",
2808 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2809 
2810 		if (db_pager_quit)
2811 			break;
2812 	}
2813 	if (count != 0)
2814 		db_printf("\n");
2815 	db_indent -= 2;
2816 }
2817 
2818 /* XXX. */
2819 #undef count
2820 
2821 /* XXX need this non-static entry for calling from vm_map_print. */
2822 void
2823 vm_object_print(
2824         /* db_expr_t */ long addr,
2825 	boolean_t have_addr,
2826 	/* db_expr_t */ long count,
2827 	char *modif)
2828 {
2829 	vm_object_print_static(addr, have_addr, count, modif);
2830 }
2831 
2832 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE)
2833 {
2834 	struct pctrie_iter pages;
2835 	vm_object_t object;
2836 	vm_page_t m, start_m;
2837 	int rcount;
2838 
2839 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2840 		db_printf("new object: %p\n", (void *)object);
2841 		if (db_pager_quit)
2842 			return;
2843 		start_m = NULL;
2844 		vm_page_iter_init(&pages, object);
2845 		VM_RADIX_FOREACH(m, &pages) {
2846 			if (start_m == NULL) {
2847 				start_m = m;
2848 				rcount = 0;
2849 			} else if (start_m->pindex + rcount != m->pindex ||
2850 			    VM_PAGE_TO_PHYS(start_m)  + ptoa(rcount) !=
2851 			    VM_PAGE_TO_PHYS(m)) {
2852 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2853 				    (long)start_m->pindex, rcount,
2854 				    (long)VM_PAGE_TO_PHYS(start_m));
2855 				if (db_pager_quit)
2856 					return;
2857 				start_m = m;
2858 				rcount = 0;
2859 			}
2860 			rcount++;
2861 		}
2862 		if (start_m != NULL) {
2863 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2864 			    (long)start_m->pindex, rcount,
2865 			    (long)VM_PAGE_TO_PHYS(start_m));
2866 			if (db_pager_quit)
2867 				return;
2868 		}
2869 	}
2870 }
2871 #endif /* DDB */
2872